D-01
Draw first.
If it cannot be specified, it cannot be trusted. Every system I ship starts as a document that someone else could build without me in the room.
§1.0 IDENTITY · DRAWING OF A PERSON
Shreyas B S
Agentic Systems Engineer
I build AI that can answer for itself.
I build AI systems the way engineers build bridges: specified completely before they are built, instrumented so they can be trusted, and shipped to survive production. Right now that means voice agents on live enterprise calls, a reputation economy for autonomous agents, and a patent-pending detector for what LLMs should never be told.
currently: ML Engineer Intern · Agentic Universe · CSE @ MIT Manipal, class of 2028 · Bengaluru
- India AI Impact Buildathon 2026
- Top 15 / 42,000+
- IISc Arbitrage Arena 2026
- 2nd Runners Up / 420+
- Patent filings
- 4
- Technoxian World Cup
- 5th / 80+
§2.0 DOCTRINE
How I think
Five rules govern everything I ship. They are not aspirations; each one is somewhere in my work, and most of them are enforced on this very page.
D-02
Determinism is respect.
Same seed, same behavior. A system that cannot be replayed cannot be debugged, audited, or believed. This site contains zero calls to Math.random(), and that is enforced at build time.
D-03
Errors are asymmetric.
Every failure mode has a price tag, and the prices are never equal. In voice AI, a false interruption is catastrophic and a missed one is recoverable, so you default to silence and escalate monotonically. Find the expensive failure and design for it.
D-04
Evals before features.
A capability you cannot measure is a liability you have not priced. I wrote 54 deterministic metrics for a voice pipeline before anyone asked for a dashboard.
D-05
Trust is infrastructure.
Reputation, provenance, and auditability are not add-ons; they are the load-bearing walls of agent systems. It is why Karma prices misbehavior, why MTSRD watches what conversations reconstruct, and why my IEEE submission is about governance that holds when some of the agents are lying.
§3.0 TRAJECTORY
The line so far
-
2024
MIT Manipal
Started Computer Science Engineering, class of 2028. Joined RoboManipal, the university's robotics team, and found the lab I would not leave.
-
2025
First production agents
Co-founded VersionTwo in June and shipped its first system: a 10-agent content workflow that cut creation time by 80% for DeepTech founders. FarmBot work at RoboManipal began winning: Technoxian World Cup, 5th of 80+.
-
EARLY 2026
The competition season
2nd Runners Up of 420+ teams at the IISc Arbitrage Arena. Top 15 of 42,000+ entries at the India AI Impact Buildathon with a multi-agent scam-baiting honeypot. Three provisional patent filings landed in twelve weeks: fuzzy control, semantic-leakage detection, slip estimation.
-
2026
Production and protection
ML Engineer Intern at Agentic Universe, building evaluation infrastructure for live enterprise voice agents. The karma reputation paper submitted to IEEE SSRR 2026, and Y Combinator Startup School India in April.
-
NEXT
The open segment
The line continues below in §4.
§4.0 SIGNAL
What I'm optimizing for
I want to work on agentic systems that have to survive production: evaluation infrastructure, safety tooling, voice agents, multi-agent orchestration. If you are building in that space and the hard parts are still hard, I want to hear about them.
open to: internships research collaborations agent infrastructure
off the clock
Robotics lab nights at RoboManipal. Teaching 250+ juniors to build agents from raw APIs, no frameworks allowed. Formula 1 on race weekends, markets the rest of the week, and an unhealthy fascination with why systems fail rather than how they work.
§5.0 TRANSMISSION
Start a conversation
This was the drawing. The machine room is on the other side.
[§1] thesis.load()
Intelligence is cheap now. Reliability isn't.
I'm Shreyas. My systems are currently evaluating live enterprise calls at Agentic Universe, scoring AI agents' trustworthiness at Versiontwo, and baiting scam callers for sport. Everything below is real, linked, and labeled with its true status. This site runs on the same rules.
currently: ML Engineer Intern · Agentic Universe · CSE @ MIT Manipal, class of 2028 · Bengaluru
- India AI Impact Buildathon 2026
- Top 15 / 42,000+
- IISc Arbitrage Arena 2026
- 2nd Runners Up / 420+
- Patent filings
- 4
- Technoxian World Cup
- 5th / 80+
[§2] experience.mount()
Experience
the pipeline, drawn: one call through four stations
80% content creation time reduction for DeepTech founders
Technoxian World Cup 5th / 80+ · Farm Robotics Challenge USA 2026
Teaching & mentoring
TDA Gen AI & Agentic AI Bootcamp · RoboManipal · Finova
I run the TDA Gen AI & Agentic AI Bootcamp at MIT Manipal for 250+ junior developers: pure APIs and Python, no frameworks allowed, because you cannot debug what you cannot see. 5+ workshops taught to 300+ people. 60+ juniors mentored inside RoboManipal. General Secretary at Finova. One node, fanned out.
[§3] machines.mount()
Projects
K = 0.35·Safety + 0.30·Accuracy + 0.20·Consistency + 0.15·Performance decay: 3% / day · critical safety failures carry a 0.6x penalty multiplier
Reputation decays unless renewed by behavior; a score you earned last quarter is not a score.
The founding simulation ran 1000+ heterogeneous agents (human, AI, and robot archetypes) in a shared economy. Altruistic strategies survived 3.36x longer than baseline. Exploiter strategies went extinct within 3 generations. Cooperation collapsed once fraudsters crossed roughly 12% of the population, and a Gini circuit-breaker cut wealth concentration by 34%.
Shipped as an open-source scoring engine: Fastify + TypeScript API, Next.js console, PostgreSQL + Redis, a Python evaluation CLI published to PyPI, and an npm score-engine package. Evaluation suites cover capability, compliance, and safety, including prompt-injection resistance.
The simulation results became the reputation-economy paper submitted to IEEE SSRR 2026 (see §4).
An INT8-quantized embedding ensemble (BGE, E5, MPNet) votes with a 2-of-3 consensus rule; matches are scored per-chunk with max coverage, never summed, and thresholds adapt per document type. Filed as an Indian provisional patent via Manipal Academy of Higher Education.
dual-LLM split: deterministic classification apart from the adaptive persona · sliding-window summarization keeps long cons cheap
Top 15 / 42,000+ · India AI Impact Buildathon 2026
The triage engine is rules-first: "why this band?" explanations cite the exact protocol rules that fired, from ANC danger signs to snakebite response. AI is routed by cost and stakes: small model for extraction, large model only for high-risk referral language.
ANC · IMNCI · HBNC · PPH · TB DOTS · snakebite · NCD PEN · IDSP
interactive frontier: illustrative · headline numbers: real 2018-2024 walk-forward backtest
6.3-year backtest · 12-asset universe · volatility 17.45% vs 21.35%
2nd Runners Up / 420+ · IISc Arbitrage Arena 2026
Dual-timescale design: signals form on 30-minute bars, execution simulates on 5-second data, roughly 450 million bars over 2011-2024. 71.2% of months finished positive. Most FX momentum is just the dollar moving; removing it first is the whole trade.
common factor out · residual signal in
The research behind the temporal-GNN paper in §4: build multiplex networks over India's top 50 companies, then learn how negative news and volatility shocks travel through them.
code for the §4 temporal-GNN paper
The same thesis as everything else here: decisions you can audit beat decisions you can only trust. Every classification cites its evidence.
every decision → span → token → rule
recall quality
next review: 6d · EF: 2.50
agents choose the content · SM-2 chooses the time · the q buttons run the real EF update
[§4] research.verify()
Research & Patents
- Karma-based reputation economies for heterogeneous human-AI-robot systems. IEEE SSRR 2026.
- IoT water quality monitoring, lean 5HO. India International WASH Conference 2026.
- Temporal GNN analysis of financial networks.
- MCDM siting of solar parks.
Patent filings
-
FILED · 202641000642
A System and Method for Detecting and Preventing Semantic Data Leakage Across Conversational Systems
-
FILED · 202541126466
A System and a Method for Real-Time Control Using Attentive Fuzzy Logic with Variable Rule and Membership Structure
-
FILED · 202641011673
Slip Estimation on SSMR and Differential Drive Robots Using IMU and Encoders for Accurate Feedback in a Closed Loop System
-
Title pending owner confirmation
Fourth filing
[§5] schematic.energize()
One system
All subsystems report to the same core. Watch the signals: multi-agent systems, evaluation, trust.
Systems that explain themselves.
That was the machine room. The person who drew it is on the other side.